Handwritten Style Recognition for Chinese Characters on HCL2020 Dataset

Structural features of Chinese characters provide abundant style information for handwritten style recognition, while prior work on this task has few senses of using structural information. Meanwhile, based on current handwritten Chinese character datasets, it is hard to obtain a good generalization model only by character category and writer information. Therefore, we add the structural information known as morpheme which is the smallest and unique structure in Chinese character into the large handwritten dataset HCL2000 and update it to HCL2020. We also present a deep fusion network (Morpheme-based Handwritten Style Recognition Network, M-HSRNet), capturing both overall layout characteristics and detail structural features of characters to recognize handwritten style. The evaluation results of the proposed model on HCL2020 are observed to prove the effectiveness of morpheme. Together with the proposed Morpheme Encoder module, our approach achieves an accuracy of 78.06% in handwritten style recognition, which is 3 points higher than the result without morpheme information.